REPAIR Approach for Social-based City Reconstruction Planning in case of natural disasters

AI summary

Overview

Research area: AI for disaster recovery planning — applying deep reinforcement learning to city reconstruction after natural disasters (arXiv:2510.19048v1 [cs.CY], published 21 Oct 2025, licensed CC BY-NC-SA 4.0).

Technical level: Intermediate. The paper assumes familiarity with reinforcement learning basics (states, actions, rewards, Q-values), graph representations of cities, and multi-objective constraint problems.

Scope: This paper describes REPAIR (post disaster REbuilding plAn ProvIdeR), a decision-support system that uses reinforcement learning to generate alternative post-disaster reconstruction plans that maximize social benefit under budget, time, political-priority and physical-dependency constraints, validated on the city of L'Aquila after its 2009 earthquake.

What This Paper Is About

After a major disaster such as an earthquake, governments must rebuild damaged buildings, roads and bridges with limited budget and time, while balancing political priorities and the needs of citizens. The authors argue that existing approaches either ignore social benefit, ignore physical dependencies such as a collapsed bridge blocking access to a hospital, or fail to model political strategy. REPAIR is presented as a generic approach that produces a set of alternative reconstruction plans for local administrators to choose from, rather than a single "best" plan.

Key Contributions

1. A formal optimization model of post-disaster reconstruction that combines budget and time limits, a political-priority constraint, a physical-dependency constraint (modelled as a directed graph), and an objective function that maximizes social benefit across the reconstruction timeline.

2. Use of Double Deep Q-Learning (DDQN) to solve the model, described as the first application of Double Deep Q-Learning to post-disaster reconstruction planning, with a customized immediate social reward function and an explicit constraints-verification procedure run at each reconstruction cycle.

3. A comparative analysis of reinforcement learning techniques, extending the authors' previously published work [14]: DDQN, Q-Learning, SARSA and Deep SARSA are compared against each other, plus a random agent used as a baseline.

4. A real-world case study and open implementation: application of REPAIR to the reconstruction of L'Aquila (Italy) after the 2009 earthquake, using shapefiles from an open online repository plus additional shapefiles provided by urban planners, with a Python implementation released publicly on Zenodo [15].

Main Findings

• The plan ordering drives social benefit. With the same set of units (hospital, school, cinema) reconstructed in different orders, the example plan P1 yields a social benefit of 11400 while plan P2 yields 9600, so P1 is the more beneficial plan. The example fixes T_e = 6 years, T_h = 2 years, T_s = 1.5 years, T_c = 1 year, S(h) = 2000, S(s) = 1000, and 600 people for the cinema unit as written in the paper.

• Social benefit is defined as people, not money. S(v) is the number of people taking direct and indirect advantage after reconstruction of unit v; direct beneficiaries are captured by b_v, and indirect beneficiaries are people living or working in the neighbourhood normalized by distance from v, weighted by constants alpha and beta with alpha + beta = 1.

• Political strategy is encoded numerically. Each unit has a political priority p_v in {1, ..., 10}, where 1 is Low Priority and 10 is Highest Priority. A plan must satisfy an average political priority threshold; in the evaluation the threshold Th_p is fixed to 8 for the first cycle of reconstruction plans, corresponding to the 80% political strategy requirement. The threshold is reduced in every cycle because less beneficial buildings are considered in later cycles.

• Physical dependencies force reconstruction order. Dependencies are modelled as a directed graph G' = (V', E') where V' is a subset of V, distinct from the undirected units graph G. If a unit is in the plan, all units connected to it in the dependency graph must also be included, and any damaged accessing road or bridge must be reconstructed first; the social benefit of the reconstructed dependency is added to the next reconstructed unit.

• REPAIR is claimed to be the first to apply DDQN here, and to be generic. The paper states that, to the best of the authors' knowledge, REPAIR is the first to apply the Double Deep Q-Learning approach to this problem, and that no prior work used reinforcement learning at all for post-disaster reconstruction planning. The approach is described as applicable to areas of any extension and able to generate alternative plans by changing input parameters (political priority, considered area, social benefit model).

• Results of the algorithm comparison are not reported in the available content. The paper states that this extension adds a comparison of DDQN, Q-Learning, SARSA and Deep SARSA, plus a random-agent baseline, and that the results on the L'Aquila case study demonstrate the soundness of the approach. However, the truncated content provided does not include the evaluation section, so no accuracy, convergence, runtime or plan-quality figures for that comparison are reported here.

• Data pipeline from shapefiles. The graph G, unit types, number of people using units, and a vulnerability index ranging from 0 to 5 (0 is the most vulnerable building, 5 the least vulnerable) are extracted from shapefiles. Unit status s_v is set to 0 if vulnerability is between 0 and 3 and 1 otherwise; reconstruction cost and time follow the approach of M. Polese et al. [41]; political priority is derived from building type [42]; the physical dependency graph G' is derived from the obtained unit statuses. Damaged building nodes are shown in red and fine buildings in green; damaged roads as red dotted edges and undamaged roads as solid black edges.

Methodology in Plain English

The authors first build two data structures for the damaged city: an undirected graph where nodes are reconstruction units (buildings, roads, bridges) and edges are connections between them, and a directed graph capturing which units cannot be rebuilt before others. They populate these from geographical shapefiles, deriving each unit's damage status from its vulnerability index, and estimating reconstruction cost and time from an established procedure.

They then write the reconstruction problem as a mathematical optimization problem: choose an ordered list of units to rebuild that maximizes total social benefit, subject to not exceeding the budget, not exceeding the total reconstruction time T_e, satisfying the political-priority threshold, and respecting the physical dependencies.

Because solving this with classical solvers is not the route the authors take, they instead train a reinforcement learning agent. The agent's state is a tuple of current location, remaining budget and remaining time; its action is adding a specific unit to the plan; its reward is the social benefit the newly reconstructed unit delivers. The agent uses Double Deep Q-Learning with convolutional neural networks to approximate the Q-function, updating Q-values with a Bellman-style equation and minimizing the squared difference between the Q-target and the Q-network. A separate verification routine checks the budget, time, political-priority and dependency constraints every time the agent adds a unit and at each reconstruction cycle, running across multiple training cycles until the budget and time are exhausted. Finally, the scheme is exercised on the historical centre of L'Aquila, damaged in the 2009 earthquake, with shapefiles from an open Abruzzo regional repository supplemented by urban-planner data.

Why This Matters

Impact on research: The paper pushes disaster-recovery planning away from purely cost- or damage-driven optimization toward plans scored by human social benefit, and it treats reconstruction as a sequential decision problem suitable for reinforcement learning rather than as a one-shot selection problem. It also introduces a comparison protocol across Q-Learning, Deep Q-Learning variants and a random baseline for this domain, and releases code for reuse.

Real-world applications:

• Municipal and regional reconstruction offices allocating limited post-earthquake budgets across buildings, roads and bridges.
• Urban planners needing to justify why a bridge or access road must be rebuilt before the hospital or school it serves.
• Policy-makers who want to test alternative political strategies (for example, prioritizing public, health and education services) before committing funds.
• Disaster-management agencies that need a set of alternative plans rather than a single recommendation, so that political and practical trade-offs can be negotiated.

Industry relevance: The approach fits the broader decision-support and smart-city software market, where geographic information systems, shapefile data and machine-learning planning tools are increasingly combined. The released Python implementation lowers the barrier for consultancies and public agencies to experiment with their own city data.

Future Directions

• Report and analyze the full comparison of DDQN, Q-Learning, SARSA and Deep SARSA against the random-agent baseline, identifying which learning algorithm is most effective and efficient (RQ6 in the paper).
• Validate REPAIR on additional real case studies and on larger or more complex areas, since the paper claims genericity and applicability to areas of any extension.
• Address the threats to validity the authors identify in Section 7 of the paper, which are not included in the available content.
• Extend the model so that adaptation to different and more complex situations — the stated reason for choosing reinforcement learning over classical solvers — is demonstrated in practice.

Target Audience

Researchers and graduate students in AI for social good, reinforcement learning and computational sustainability; civil engineering, urban planning and disaster-management researchers; and public-sector or consultancy analysts who build decision-support tools for post-disaster recovery. Readers without any reinforcement learning background will need to read the state/action/reward definitions carefully, but the key concepts of social benefit, political priority and physical dependency are explained in plain terms.

Authors’ abstract

Natural disasters always have several effects on human lives. It is challenging for governments to tackle these incidents and to rebuild the economic, social and physical infrastructures and facilities with the available resources (mainly budget and time). Governments always define plans and policies according to the law and political strategies that should maximise social benefits. The severity of damage and the vast resources needed to bring life back to normality make such reconstruction a challenge. This article is the extension of our previously published work by conducting comprehensive comparative analysis by integrating additional deep learning models plus random agent which is used as a baseline. Our prior research introduced a decision support system by using the Deep Reinforcement Learning technique for the planning of post-disaster city reconstruction, maximizing the social benefit of the reconstruction process, considering available resources, meeting the needs of the broad community stakeholders (like citizens' social benefits and politicians' priorities) and keeping in consideration city's structural constraints (like dependencies among roads and buildings). The proposed approach, named post disaster REbuilding plAn ProvIdeR (REPAIR) is generic. It can determine a set of alternative plans for local administrators who select the ideal one to implement, and it can be applied to areas of any extension. We show the application of REPAIR in a real use case, i.e., to the L'Aquila reconstruction process, damaged in 2009 by a major earthquake.